The Short-Term Prediction of Length of Day Using 1D Convolutional Neural Networks (1D CNN)

Area of Science:

• Geophysics
• Astrodynamics
• Machine Learning

Background:

• Accurate Earth orientation parameter (EOP) predictions are vital for space mission tracking and navigation.
• Forecasting length of day (LOD) variations is challenging due to atmospheric circulation torques.

Purpose of the Study:

• To develop a novel time-series prediction algorithm for Earth's length of day (LOD).
• To enhance LOD prediction accuracy by incorporating atmospheric angular momentum (AAM) data.

Main Methods:

• Utilized a one-dimensional convolutional neural network (1D CNN), a deep learning approach.
• Applied least squares (LS) method for detrending LOD and Z-component AAM series.
• Integrated IERS EOP 14 C04 data and GFZ's AAM axial Z component for prediction.

Main Results:

• The 1D CNN model successfully predicted LOD variations.
• Incorporating AAM data significantly improved prediction accuracy.
• The developed algorithm demonstrated potential for reconstructing and predicting LOD up to 7 days.

Conclusions:

• The proposed 1D CNN method offers an optimal solution for LOD prediction.
• This approach enhances the reliability of EOP predictions for critical applications.
• Accurate LOD forecasting is achievable with advanced machine learning techniques and relevant geophysical data.